# Smoke config: pg_pinecone_native plumbing proof, not a comparison run of
# record. Same dataset shape as configs/smoke.yaml (identical seed/n/dim/
# n_centers/cluster_sigma/cluster_scale_mix/background_frac -- reuses the
# same cached synthetic dataset + hardness calibration) so pgvector_hnsw's
# numbers here are directly comparable to the committed smoke run, but the
# system list, filter matrix, and concurrency are trimmed hard for wall time:
# this config exists to prove the harness/engine PLUMBING end to end
# (EXPLAIN-gated index usage, pinecone.last_scan_mode() assertions, tie-aware
# recall, plan-shape tags, provenance stamps) at 50k rows, not to produce a
# result worth reporting as a benchmark number. Run with
# `--data-dir` pointed at a scratch directory on a free port -- NEVER the
# default bench/vector/.pgdata while another run (e.g. a baselines series) might
# already own it.
name: smoke_pg_pinecone_native
seed: 42
k: 10
warmup_queries: 10
hardness_recall_threshold: 0.97
min_throughput_seconds: 1.5

dataset:
  kind: synthetic
  n: 50000
  dim: 768
  n_centers: 1000
  cluster_sigma: 1.1
  cluster_scale_mix: [0.6, 1.0, 1.6]
  background_frac: 0.02
  n_queries: 500

metric: l2

filters:
  selectivities: [0.01]
  modes: [correlated]
  run:
    - {selectivity: 0.01, mode: correlated}

systems:
  # Comparison anchor -- trimmed sweep (2 points, 1 conn only) purely so a
  # plumbing run finishes in minutes; configs/smoke.yaml is the real
  # pgvector_hnsw smoke result of record.
  - name: pgvector_hnsw
    type: hnsw
    build_params: {m: 16, ef_construction: 64}
    sweep_param: ef_search
    sweep_values: [10, 40]
    iterative_scan_values: [null]
    filtered_iterative_scan_values: ["relaxed_order"]
    scan_budget_param: hnsw.max_scan_tuples
    filtered_scan_budget_values: [auto]

  # Unfiltered build (no filter_column) -- exercises the plain ivf_scan path
  # (MODE_H expected for every unfiltered query; n=50000 > small_n_threshold
  # so this is NOT the small-N flat bootstrap).
  - name: pg_pinecone_native_unfilt
    type: pg_pinecone_native
    build_params: {seed: 42, k_bits: 512, small_n_threshold: 1000, rescore_multiplier: 10}
    sweep_param: rescore_multiplier
    sweep_values: [4, 10, 20]

  # Filtered build against the 1% correlated filter's equality-mirror column
  # (loader.filter_eq_column_name), auto mode selection: realized selectivity
  # ~0.01 <= the default pinecone.mode_l_threshold (0.03), so mode L is expected
  # -- filter-first exact scan, recall 1.0 by construction.
  - name: pg_pinecone_native_filt_l
    type: pg_pinecone_native
    build_params:
      seed: 42
      k_bits: 512
      small_n_threshold: 1000
      rescore_multiplier: 10
      mode_l_threshold: 0.03
      filter_column: filt_correlated_0010_eq
    sweep_param: rescore_multiplier
    sweep_values: [4, 10, 20]

  # Same build, but pinecone.force_scan_mode='h' forces the IVF-probe + in-loop
  # filter path regardless of estimated selectivity -- exercises the OTHER
  # mode-assertion branch (mode H) that the mode_l_threshold default alone
  # wouldn't reach at this selectivity.
  - name: pg_pinecone_native_filt_h
    type: pg_pinecone_native
    build_params:
      seed: 42
      k_bits: 512
      small_n_threshold: 1000
      rescore_multiplier: 10
      filter_column: filt_correlated_0010_eq
      force_scan_mode: h
    sweep_param: rescore_multiplier
    sweep_values: [4, 10, 20]

concurrency: [1]

churn:
  enabled: false

sift1m:
  enabled: false
